The following inequality is given in some of Yale's online lecture notes

$$P(|Z|>x) \leq 2 \sqrt{2 \pi} \phi(x)$$

Where $Z \sim N(0,1)$ with density $\phi(x)$. They call it a concentration inequality, the proof is not given.

I know how to prove the bounds

$$(1/x - 1/x^3)\phi(x) \leq 1-\Phi(x) \leq 1/x \phi(x)$$

But the method I used for these did not get me anywhere.

I know that it is enough to show

$$\Phi(x) \geq 1- \sqrt{2 \pi} \phi(x)$$

by symmetry of normal distribution. So we need to show

$$\int_x^\infty \phi(y)dy \geq 1- \sqrt{2 \pi} \phi(x)$$


Following advice from answer

Note that $E[e^{Zt}] = e^{t^2/2}$ and we already have the inequality

$$P(Z>x) \leq E[e^{Zt}]/e^{xt}$$

Maximising the RHS equality with respect to $t$ gives $t_\max = x$, and gives the upperbound needed.


It is not true. When $x=0.1$, $P(|Z|>x) = 0.92$ and $2\phi(x)=0.79$.

Maybe the correct one is:

$$P(|Z|>x) \leq 2 \sqrt{2\pi}\phi(x)$$

For new inequality, following follow step:

  1. Get $E(e^{\lambda Z})$

  2. Find $\sup_\lambda(\lambda x - \log E(e^{\lambda Z}))$

  3. Use Chernoff bound, you can get $\Pr(Z>x) < e^{-t^2/2}$

| cite | improve this answer | |
  • $\begingroup$ You're right. I forgot the normalisingn constant is not included. Will correct my post $\endgroup$ – Xiaomi Dec 1 '18 at 7:46
  • $\begingroup$ Thanks a lot! Maximising the bound never occured to me but it's so obvious now :) $\endgroup$ – Xiaomi Dec 1 '18 at 11:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.